Continue 开源 IDE 编程助手

当“AI 原生编辑器”开始成为一个品类，开发者面前似乎出现了整套迁移题：要不要换掉现在的编辑器？扩展、快捷键和团队配置怎么办？代码是否必须交给某一家模型服务？Continue 给出的回答没有那么戏剧化——先装一个侧栏。

2023 年，Continue 以开源的 VS Code 与 JetBrains 扩展发展，把聊天、补全和代码库上下文带进开发者已经使用的 IDE。它没有要求所有人接受同一模型。云端 API、本地模型和不同提供商可以按需要接入，工作流也能被修改。编辑器留下，模型成为可替换的部件。这种选择的价值不只在价格。企业可以审计扩展如何处理上下文，团队可以决定哪些代码发送到哪里，也可以在能力、成本与数据边界之间做自己的取舍。一个模型在补全上更快，另一个在代码库问答上更好，并不需要连编辑器一起更换。开源在这里不是抽象口号，而是一种配置权。

配置权也有成本。模型越可换，团队越要理解不同后端的上下文窗口、接口差异和部署要求。本地模型减少了外发数据，却可能带来速度和质量折衷；自托管提高控制，也把运维责任带回来。Continue 提供的是可选择的结构，不是自动获得最佳结果的按钮。

它与 Cursor、Windsurf 代表的完整编辑器路线形成了清晰对照。一体化产品可以同时优化索引、界面与模型调用，交互更顺；开放侧栏则允许开发者保留工具链，把关键部件拆开选择。两者争夺的并非同一种自由：前者减少操作摩擦，后者减少退出与替换的摩擦。

Continue 的位置因此不需要靠“最强 Agent”来证明。它让 AI 编程生态多出一条岔路：当行业不断要求开发者进入新的封闭工作台时，也有人坚持把模型带到已有工作台里，并把门的钥匙留在使用者手上。

As the “AI-native editor” became a category, developers appeared to face a full migration problem. Should they leave their current editor? What would happen to extensions, shortcuts, and team settings? Did using AI require sending code to one model provider? Continue offered a less dramatic answer: begin with a sidebar.

During 2023, Continue developed as an open-source extension for VS Code and JetBrains, bringing chat, completion, and repository context into IDEs developers already used. It did not require everyone to accept the same model. Cloud APIs, local models, and different providers could be connected as needed, and workflows could be customized. The editor stayed; the model became a replaceable component. The value of that choice was not only price. An organization could inspect how the extension handled context, decide where code was sent, and make its own trade-offs among capability, cost, and data boundaries. If one model was faster at completion and another stronger at repository questions, replacing the model did not require replacing the editor. Open source here meant configuration authority rather than a general slogan.

Authority came with work. The more interchangeable the models, the more a team had to understand differences in context windows, interfaces, and deployment. A local model could reduce data leaving the machine while trading away speed or quality. Self-hosting increased control and returned operational responsibility. Continue offered a structure for choosing, not a button that guaranteed the best result.

That structure formed a useful contrast with the complete-editor route represented by Cursor and Windsurf. An integrated product could optimize indexing, interface, and model calls together, reducing interaction friction. An open sidebar let developers keep their toolchain and choose critical parts separately. The two routes pursued different kinds of freedom: one reduced the friction of operating; the other reduced the friction of leaving and replacing.

Continue's place does not depend on having been the “strongest agent.” It added a fork to the AI-coding ecosystem. While the industry repeatedly invited developers into new enclosed workspaces, an open extension insisted that the model could come to the existing workspace—and that the user could keep the key to the door.